Self-improving RAG pipeline benchmarked on Meta's CRAG — autonomous optimization loop, hybrid retrieval, hallucination controls. +17% accuracy, -46% hallucinations.

Transform vague optimization problems into fully scaffolded autonomous experiment loops. Claude Code skill.

Autonomous training optimizer for nanoGPT-style models with governed patch search, empirical validation, and rollback-safe execution.

Benchmark-driven autonomous optimization for bounded systems